lustre/ldlm/ldlm_flock.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Hewlett-Packard Development Company LP.
   5  *   Developed under the sponsorship of the US Government under
   6  *   Subcontract No. B514193
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 #include <lustre_dlm.h>
  31 #include <obd_support.h>
  32 #include <obd_class.h>
  33 #include <lustre_lib.h>
  34 #include <libcfs/list.h>
  35 #else
  36 #include <liblustre.h>
  37 #include <obd_class.h>
  38 #endif
  39
  40 #include "ldlm_internal.h"
  41
  42 #define l_flock_waitq   l_lru
  43
  44 static struct list_head ldlm_flock_waitq = CFS_LIST_HEAD_INIT(ldlm_flock_waitq);
  45
  46 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
  47                             void *data, int flag);
  48
  49 /**
  50  * list_for_remaining_safe - iterate over the remaining entries in a list
  51  *              and safeguard against removal of a list entry.
  52  * @pos:        the &struct list_head to use as a loop counter. pos MUST
  53  *              have been initialized prior to using it in this macro.
  54  * @n:          another &struct list_head to use as temporary storage
  55  * @head:       the head for your list.
  56  */
  57 #define list_for_remaining_safe(pos, n, head) \
  58         for (n = pos->next; pos != (head); pos = n, n = pos->next)
  59
  60 static inline int
  61 ldlm_same_flock_owner(struct ldlm_lock *lock, struct ldlm_lock *new)
  62 {
  63         return((new->l_policy_data.l_flock.pid ==
  64                 lock->l_policy_data.l_flock.pid) &&
  65                (new->l_export == lock->l_export));
  66 }
  67
  68 static inline int
  69 ldlm_flocks_overlap(struct ldlm_lock *lock, struct ldlm_lock *new)
  70 {
  71         return((new->l_policy_data.l_flock.start <=
  72                 lock->l_policy_data.l_flock.end) &&
  73                (new->l_policy_data.l_flock.end >=
  74                 lock->l_policy_data.l_flock.start));
  75 }
  76
  77 static inline void
  78 ldlm_flock_destroy(struct ldlm_lock *lock, ldlm_mode_t mode, int flags)
  79 {
  80         ENTRY;
  81
  82         LDLM_DEBUG(lock, "ldlm_flock_destroy(mode: %d, flags: 0x%x)",
  83                    mode, flags);
  84
  85         LASSERT(list_empty(&lock->l_flock_waitq));
  86
  87         list_del_init(&lock->l_res_link);
  88         if (flags == LDLM_FL_WAIT_NOREPROC) {
  89                 /* client side - set a flag to prevent sending a CANCEL */
  90                 lock->l_flags |= LDLM_FL_LOCAL_ONLY | LDLM_FL_CBPENDING;
  91
  92                 /* when reaching here, it is under lock_res_and_lock(). Thus,
  93                    need call the nolock version of ldlm_lock_decref_internal*/
  94                 ldlm_lock_decref_internal_nolock(lock, mode);
  95         }
  96
  97         ldlm_lock_destroy_nolock(lock);
  98         EXIT;
  99 }
 100
 101 static int
 102 ldlm_flock_deadlock(struct ldlm_lock *req, struct ldlm_lock *blocking_lock)
 103 {
 104         struct obd_export *req_export = req->l_export;
 105         struct obd_export *blocking_export = blocking_lock->l_export;
 106         pid_t req_pid = req->l_policy_data.l_flock.pid;
 107         pid_t blocking_pid = blocking_lock->l_policy_data.l_flock.pid;
 108         struct ldlm_lock *lock;
 109
 110 restart:
 111         list_for_each_entry(lock, &ldlm_flock_waitq, l_flock_waitq) {
 112                 if ((lock->l_policy_data.l_flock.pid != blocking_pid) ||
 113                     (lock->l_export != blocking_export))
 114                         continue;
 115
 116                 blocking_pid = lock->l_policy_data.l_flock.blocking_pid;
 117                 blocking_export = (struct obd_export *)(long)
 118                         lock->l_policy_data.l_flock.blocking_export;
 119                 if (blocking_pid == req_pid && blocking_export == req_export)
 120                         return 1;
 121
 122                 goto restart;
 123         }
 124
 125         return 0;
 126 }
 127
 128 int
 129 ldlm_process_flock_lock(struct ldlm_lock *req, int *flags, int first_enq,
 130                         ldlm_error_t *err, struct list_head *work_list)
 131 {
 132         struct ldlm_resource *res = req->l_resource;
 133         struct ldlm_namespace *ns = res->lr_namespace;
 134         struct list_head *tmp;
 135         struct list_head *ownlocks = NULL;
 136         struct ldlm_lock *lock = NULL;
 137         struct ldlm_lock *new = req;
 138         struct ldlm_lock *new2 = NULL;
 139         ldlm_mode_t mode = req->l_req_mode;
 140         int local = ns_is_client(ns);
 141         int added = (mode == LCK_NL);
 142         int overlaps = 0;
 143         int splitted = 0;
 144         ENTRY;
 145
 146         CDEBUG(D_DLMTRACE, "flags %#x pid %u mode %u start "LPU64" end "LPU64
 147                "\n", *flags, new->l_policy_data.l_flock.pid, mode,
 148                req->l_policy_data.l_flock.start,
 149                req->l_policy_data.l_flock.end);
 150
 151         *err = ELDLM_OK;
 152
 153         if (local) {
 154                 /* No blocking ASTs are sent to the clients for
 155                  * Posix file & record locks */
 156                 req->l_blocking_ast = NULL;
 157         } else {
 158                 /* Called on the server for lock cancels. */
 159                 req->l_blocking_ast = ldlm_flock_blocking_ast;
 160         }
 161
 162 reprocess:
 163         if ((*flags == LDLM_FL_WAIT_NOREPROC) || (mode == LCK_NL)) {
 164                 /* This loop determines where this processes locks start
 165                  * in the resource lr_granted list. */
 166                 list_for_each(tmp, &res->lr_granted) {
 167                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 168                         if (ldlm_same_flock_owner(lock, req)) {
 169                                 ownlocks = tmp;
 170                                 break;
 171                         }
 172                 }
 173         } else {
 174                 lockmode_verify(mode);
 175
 176                 /* This loop determines if there are existing locks
 177                  * that conflict with the new lock request. */
 178                 list_for_each(tmp, &res->lr_granted) {
 179                         lock = list_entry(tmp, struct ldlm_lock, l_res_link);
 180
 181                         if (ldlm_same_flock_owner(lock, req)) {
 182                                 if (!ownlocks)
 183                                         ownlocks = tmp;
 184                                 continue;
 185                         }
 186
 187                         /* locks are compatible, overlap doesn't matter */
 188                         if (lockmode_compat(lock->l_granted_mode, mode))
 189                                 continue;
 190
 191                         if (!ldlm_flocks_overlap(lock, req))
 192                                 continue;
 193
 194                         if (!first_enq)
 195                                 RETURN(LDLM_ITER_CONTINUE);
 196
 197                         if (*flags & LDLM_FL_BLOCK_NOWAIT) {
 198                                 ldlm_flock_destroy(req, mode, *flags);
 199                                 *err = -EAGAIN;
 200                                 RETURN(LDLM_ITER_STOP);
 201                         }
 202
 203                         if (*flags & LDLM_FL_TEST_LOCK) {
 204                                 ldlm_flock_destroy(req, mode, *flags);
 205                                 req->l_req_mode = lock->l_granted_mode;
 206                                 req->l_policy_data.l_flock.pid =
 207                                         lock->l_policy_data.l_flock.pid;
 208                                 req->l_policy_data.l_flock.start =
 209                                         lock->l_policy_data.l_flock.start;
 210                                 req->l_policy_data.l_flock.end =
 211                                         lock->l_policy_data.l_flock.end;
 212                                 *flags |= LDLM_FL_LOCK_CHANGED;
 213                                 RETURN(LDLM_ITER_STOP);
 214                         }
 215
 216                         if (ldlm_flock_deadlock(req, lock)) {
 217                                 ldlm_flock_destroy(req, mode, *flags);
 218                                 *err = -EDEADLK;
 219                                 RETURN(LDLM_ITER_STOP);
 220                         }
 221
 222                         req->l_policy_data.l_flock.blocking_pid =
 223                                 lock->l_policy_data.l_flock.pid;
 224                         req->l_policy_data.l_flock.blocking_export =
 225                                 (long)(void *)lock->l_export;
 226
 227                         LASSERT(list_empty(&req->l_flock_waitq));
 228                         list_add_tail(&req->l_flock_waitq, &ldlm_flock_waitq);
 229
 230                         ldlm_resource_add_lock(res, &res->lr_waiting, req);
 231                         *flags |= LDLM_FL_BLOCK_GRANTED;
 232                         RETURN(LDLM_ITER_STOP);
 233                 }
 234         }
 235
 236         if (*flags & LDLM_FL_TEST_LOCK) {
 237                 ldlm_flock_destroy(req, mode, *flags);
 238                 req->l_req_mode = LCK_NL;
 239                 *flags |= LDLM_FL_LOCK_CHANGED;
 240                 RETURN(LDLM_ITER_STOP);
 241         }
 242
 243         /* In case we had slept on this lock request take it off of the
 244          * deadlock detection waitq. */
 245         list_del_init(&req->l_flock_waitq);
 246
 247         /* Scan the locks owned by this process that overlap this request.
 248          * We may have to merge or split existing locks. */
 249
 250         if (!ownlocks)
 251                 ownlocks = &res->lr_granted;
 252
 253         list_for_remaining_safe(ownlocks, tmp, &res->lr_granted) {
 254                 lock = list_entry(ownlocks, struct ldlm_lock, l_res_link);
 255
 256                 if (!ldlm_same_flock_owner(lock, new))
 257                         break;
 258
 259                 if (lock->l_granted_mode == mode) {
 260                         /* If the modes are the same then we need to process
 261                          * locks that overlap OR adjoin the new lock. The extra
 262                          * logic condition is necessary to deal with arithmetic
 263                          * overflow and underflow. */
 264                         if ((new->l_policy_data.l_flock.start >
 265                              (lock->l_policy_data.l_flock.end + 1))
 266                             && (lock->l_policy_data.l_flock.end !=
 267                                 OBD_OBJECT_EOF))
 268                                 continue;
 269
 270                         if ((new->l_policy_data.l_flock.end <
 271                              (lock->l_policy_data.l_flock.start - 1))
 272                             && (lock->l_policy_data.l_flock.start != 0))
 273                                 break;
 274
 275                         if (new->l_policy_data.l_flock.start <
 276                             lock->l_policy_data.l_flock.start) {
 277                                 lock->l_policy_data.l_flock.start =
 278                                         new->l_policy_data.l_flock.start;
 279                         } else {
 280                                 new->l_policy_data.l_flock.start =
 281                                         lock->l_policy_data.l_flock.start;
 282                         }
 283
 284                         if (new->l_policy_data.l_flock.end >
 285                             lock->l_policy_data.l_flock.end) {
 286                                 lock->l_policy_data.l_flock.end =
 287                                         new->l_policy_data.l_flock.end;
 288                         } else {
 289                                 new->l_policy_data.l_flock.end =
 290                                         lock->l_policy_data.l_flock.end;
 291                         }
 292
 293                         if (added) {
 294                                 ldlm_flock_destroy(lock, mode, *flags);
 295                         } else {
 296                                 new = lock;
 297                                 added = 1;
 298                         }
 299                         continue;
 300                 }
 301
 302                 if (new->l_policy_data.l_flock.start >
 303                     lock->l_policy_data.l_flock.end)
 304                         continue;
 305
 306                 if (new->l_policy_data.l_flock.end <
 307                     lock->l_policy_data.l_flock.start)
 308                         break;
 309
 310                 ++overlaps;
 311
 312                 if (new->l_policy_data.l_flock.start <=
 313                     lock->l_policy_data.l_flock.start) {
 314                         if (new->l_policy_data.l_flock.end <
 315                             lock->l_policy_data.l_flock.end) {
 316                                 lock->l_policy_data.l_flock.start =
 317                                         new->l_policy_data.l_flock.end + 1;
 318                                 break;
 319                         }
 320                         ldlm_flock_destroy(lock, lock->l_req_mode, *flags);
 321                         continue;
 322                 }
 323                 if (new->l_policy_data.l_flock.end >=
 324                     lock->l_policy_data.l_flock.end) {
 325                         lock->l_policy_data.l_flock.end =
 326                                 new->l_policy_data.l_flock.start - 1;
 327                         continue;
 328                 }
 329
 330                 /* split the existing lock into two locks */
 331
 332                 /* if this is an F_UNLCK operation then we could avoid
 333                  * allocating a new lock and use the req lock passed in
 334                  * with the request but this would complicate the reply
 335                  * processing since updates to req get reflected in the
 336                  * reply. The client side replays the lock request so
 337                  * it must see the original lock data in the reply. */
 338
 339                 /* XXX - if ldlm_lock_new() can sleep we should
 340                  * release the ns_lock, allocate the new lock,
 341                  * and restart processing this lock. */
 342                 if (!new2) {
 343                         unlock_res_and_lock(req);
 344                          new2 = ldlm_lock_create(ns, res->lr_name, LDLM_FLOCK,
 345                                         lock->l_granted_mode, NULL, NULL, NULL,
 346                                         NULL, 0);
 347                         lock_res_and_lock(req);
 348                         if (!new2) {
 349                                 ldlm_flock_destroy(req, lock->l_granted_mode, *flags);
 350                                 *err = -ENOLCK;
 351                                 RETURN(LDLM_ITER_STOP);
 352                         }
 353                         goto reprocess;
 354                 }
 355
 356                 splitted = 1;
 357
 358                 new2->l_granted_mode = lock->l_granted_mode;
 359                 new2->l_policy_data.l_flock.pid =
 360                         new->l_policy_data.l_flock.pid;
 361                 new2->l_policy_data.l_flock.start =
 362                         lock->l_policy_data.l_flock.start;
 363                 new2->l_policy_data.l_flock.end =
 364                         new->l_policy_data.l_flock.start - 1;
 365                 lock->l_policy_data.l_flock.start =
 366                         new->l_policy_data.l_flock.end + 1;
 367                 new2->l_conn_export = lock->l_conn_export;
 368                 if (lock->l_export != NULL) {
 369                         new2->l_export = class_export_get(lock->l_export);
 370                         spin_lock(&new2->l_export->exp_ldlm_data.led_lock);
 371                         list_add(&new2->l_export_chain,
 372                                  &new2->l_export->exp_ldlm_data.led_held_locks);
 373                         spin_unlock(&new2->l_export->exp_ldlm_data.led_lock);
 374                 }
 375                 if (*flags == LDLM_FL_WAIT_NOREPROC) {
 376                         ldlm_lock_addref_internal_nolock(new2, lock->l_granted_mode);
 377                 }
 378
 379                 /* insert new2 at lock */
 380                 ldlm_resource_add_lock(res, ownlocks, new2);
 381                 LDLM_LOCK_PUT(new2);
 382                 break;
 383         }
 384
 385         /* if new2 is created but never used, destroy it*/
 386         if (splitted == 0 && new2 != NULL)
 387                 ldlm_lock_destroy_nolock(new2);
 388
 389         /* At this point we're granting the lock request. */
 390         req->l_granted_mode = req->l_req_mode;
 391
 392         /* Add req to the granted queue before calling ldlm_reprocess_all(). */
 393         if (!added) {
 394                 list_del_init(&req->l_res_link);
 395                 /* insert new lock before ownlocks in list. */
 396                 ldlm_resource_add_lock(res, ownlocks, req);
 397         }
 398
 399         if (*flags != LDLM_FL_WAIT_NOREPROC) {
 400                 if (first_enq) {
 401                         /* If this is an unlock, reprocess the waitq and
 402                          * send completions ASTs for locks that can now be
 403                          * granted. The only problem with doing this
 404                          * reprocessing here is that the completion ASTs for
 405                          * newly granted locks will be sent before the unlock
 406                          * completion is sent. It shouldn't be an issue. Also
 407                          * note that ldlm_process_flock_lock() will recurse,
 408                          * but only once because first_enq will be false from
 409                          * ldlm_reprocess_queue. */
 410                         if ((mode == LCK_NL) && overlaps) {
 411                                 struct list_head rpc_list
 412                                                     = CFS_LIST_HEAD_INIT(rpc_list);
 413                                 int rc;
 414 restart:
 415                                 ldlm_reprocess_queue(res, &res->lr_waiting,
 416                                                      &rpc_list);
 417
 418                                 unlock_res_and_lock(req);
 419                                 rc = ldlm_run_cp_ast_work(&rpc_list);
 420                                 lock_res_and_lock(req);
 421                                 if (rc == -ERESTART)
 422                                         GOTO(restart, -ERESTART);
 423                        }
 424                 } else {
 425                         LASSERT(req->l_completion_ast);
 426                         ldlm_add_ast_work_item(req, NULL, work_list);
 427                 }
 428         }
 429
 430         /* In case we're reprocessing the requested lock we can't destroy
 431          * it until after calling ldlm_ast_work_item() above so that lawi()
 432          * can bump the reference count on req. Otherwise req could be freed
 433          * before the completion AST can be sent.  */
 434         if (added)
 435                 ldlm_flock_destroy(req, mode, *flags);
 436
 437         ldlm_resource_dump(D_OTHER, res);
 438         RETURN(LDLM_ITER_CONTINUE);
 439 }
 440
 441 struct ldlm_flock_wait_data {
 442         struct ldlm_lock *fwd_lock;
 443         int               fwd_generation;
 444 };
 445
 446 static void
 447 ldlm_flock_interrupted_wait(void *data)
 448 {
 449         struct ldlm_lock *lock;
 450         struct lustre_handle lockh;
 451         int rc;
 452         ENTRY;
 453
 454         lock = ((struct ldlm_flock_wait_data *)data)->fwd_lock;
 455
 456         /* take lock off the deadlock detection waitq. */
 457         list_del_init(&lock->l_flock_waitq);
 458
 459         /* client side - set flag to prevent lock from being put on lru list */
 460         lock->l_flags |= LDLM_FL_CBPENDING;
 461
 462         ldlm_lock_decref_internal(lock, lock->l_req_mode);
 463         ldlm_lock2handle(lock, &lockh);
 464         rc = ldlm_cli_cancel(&lockh);
 465         if (rc != ELDLM_OK)
 466                 CERROR("ldlm_cli_cancel: %d\n", rc);
 467
 468         EXIT;
 469 }
 470
 471 int
 472 ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 473 {
 474         struct ldlm_namespace *ns;
 475         cfs_flock_t *getlk = lock->l_ast_data;
 476         struct ldlm_flock_wait_data fwd;
 477         struct obd_device *obd;
 478         struct obd_import *imp = NULL;
 479         ldlm_error_t err;
 480         int rc = 0;
 481         struct l_wait_info lwi;
 482         ENTRY;
 483
 484         CDEBUG(D_DLMTRACE, "flags: 0x%x data: %p getlk: %p\n",
 485                flags, data, getlk);
 486
 487         LASSERT(flags != LDLM_FL_WAIT_NOREPROC);
 488
 489         if (!(flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
 490                        LDLM_FL_BLOCK_CONV)))
 491                 goto  granted;
 492
 493         LDLM_DEBUG(lock, "client-side enqueue returned a blocked lock, "
 494                    "sleeping");
 495
 496         fwd.fwd_lock = lock;
 497         obd = class_exp2obd(lock->l_conn_export);
 498
 499         /* if this is a local lock, then there is no import */
 500         if (obd != NULL)
 501                 imp = obd->u.cli.cl_import;
 502
 503         if (imp != NULL) {
 504                 spin_lock(&imp->imp_lock);
 505                 fwd.fwd_generation = imp->imp_generation;
 506                 spin_unlock(&imp->imp_lock);
 507         }
 508
 509         lwi = LWI_TIMEOUT_INTR(0, NULL, ldlm_flock_interrupted_wait, &fwd);
 510
 511         /* Go to sleep until the lock is granted. */
 512         rc = l_wait_event(lock->l_waitq,
 513                           ((lock->l_req_mode == lock->l_granted_mode) ||
 514                            lock->l_destroyed), &lwi);
 515
 516         LDLM_DEBUG(lock, "client-side enqueue waking up: rc = %d", rc);
 517         RETURN(rc);
 518
 519 granted:
 520
 521         LDLM_DEBUG(lock, "client-side enqueue granted");
 522         ns = lock->l_resource->lr_namespace;
 523         lock_res_and_lock(lock);
 524
 525         /* take lock off the deadlock detection waitq. */
 526         list_del_init(&lock->l_flock_waitq);
 527
 528         /* ldlm_lock_enqueue() has already placed lock on the granted list. */
 529         list_del_init(&lock->l_res_link);
 530
 531         if (flags & LDLM_FL_TEST_LOCK) {
 532                 /* fcntl(F_GETLK) request */
 533                 /* The old mode was saved in getlk->fl_type so that if the mode
 534                  * in the lock changes we can decref the approprate refcount. */
 535                 ldlm_flock_destroy(lock, cfs_flock_type(getlk), LDLM_FL_WAIT_NOREPROC);
 536                 switch (lock->l_granted_mode) {
 537                 case LCK_PR:
 538                         cfs_flock_set_type(getlk, F_RDLCK);
 539                         break;
 540                 case LCK_PW:
 541                         cfs_flock_set_type(getlk, F_WRLCK);
 542                         break;
 543                 default:
 544                         cfs_flock_set_type(getlk, F_UNLCK);
 545                 }
 546                 cfs_flock_set_pid(getlk, (pid_t)lock->l_policy_data.l_flock.pid);
 547                 cfs_flock_set_start(getlk, (loff_t)lock->l_policy_data.l_flock.start);
 548                 cfs_flock_set_end(getlk, (loff_t)lock->l_policy_data.l_flock.end);
 549         } else {
 550                 int noreproc = LDLM_FL_WAIT_NOREPROC;
 551
 552                 /* We need to reprocess the lock to do merges or splits
 553                  * with existing locks owned by this process. */
 554                 ldlm_process_flock_lock(lock, &noreproc, 1, &err, NULL);
 555                 if (flags == 0)
 556                         cfs_waitq_signal(&lock->l_waitq);
 557         }
 558         unlock_res_and_lock(lock);
 559         RETURN(0);
 560 }
 561 EXPORT_SYMBOL(ldlm_flock_completion_ast);
 562
 563 int ldlm_flock_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 564                             void *data, int flag)
 565 {
 566         struct ldlm_namespace *ns;
 567         ENTRY;
 568
 569         LASSERT(lock);
 570         LASSERT(flag == LDLM_CB_CANCELING);
 571
 572         ns = lock->l_resource->lr_namespace;
 573
 574         /* take lock off the deadlock detection waitq. */
 575         lock_res_and_lock(lock);
 576         list_del_init(&lock->l_flock_waitq);
 577         unlock_res_and_lock(lock);
 578         RETURN(0);
 579 }